First of all, apparent (strict?) correlation of phenotypic parameters implies significant overlap of the molecular machinery involved in elucidating said phenotypes. If a hypothetical kinase A, when active, is induces feature X and feature Y, then any upstream activator of A is going to result in both X and Y. Experimentally, this exhibits itself as a strict correlation of X and Y. I believe this is the rationale behind reducing the dimensions of analysis by dropping highly correlated features which add data without increasing information content.

That said, I see the question as being one of determining the amount of time one is willing to invest in elucidating the exact hierarchy of the signaling pathways, and then from that selecting a defined number of molecular markers that can be screened alongside the phenotypic features. If, say, you have a highly selective inhibitor for our hypothetical kinase A, and you co-treat with another inhibitor or activator and you rescue the phenotype, or if you decouple the correlation between X and Y, then you can deduce, based on a prior knowledge, that your co-treatment is acting downstream of A, and only affecting the pathway that leads to X OR that which leads to Y (and that’s assuming there really are sufficiently distinct pathways leading to X and Y). However, I think this kind of increased resolution requires a lot more knowledge of the individual pathways and signaling members, not to mention the exact hierarchy and temporal activities of all key regulators. Considering that we don’t yet have that kind of knowledge, and compounding that with the fact that highly selective inhibitors can still fit in just one of your hands, I suspect that this kind of analysis comes with an inherent risk of generating hopelessly entangled information.

On the other hand, what about the interesting hits you miss, not because they failed to do anything interesting inside the cells, but because they somehow neutralized their molecular effects by dually hitting both inhibitory and stimulatory pathways? In that case, we need to ask ourselves, how much return would such a find bring us? For an organic chemist, it can be a very interesting challenge to try and derivatize a compound that only hits one pathway or the other rather than both, and with some luck he might actually produce a compound that is more potent than anything out there in the library-sphere. But that takes a lot of time and might not be worth the time in a scouting approach.

I don’t have an exact idea on how to find a happy middle ground, but I do think there is some merit in finding a middle ground rather than confining to either too high or too low resolution. Perhaps a small number of molecular markers can be screened with positive and negative controls and statistically ranked based on correlation with desirable phenotypes? If one or two highly correlated molecular markers are discovered, then maybe these can be used to add depth to the analysis without introducing too many dimensions or diluting the analysis beyond usefulness.

We don’t need to move away from our comfort zone. That is, we don’t need to work with non-neuronal cell types and other channels than nuclei and BIIITubulin. Certainly, it would be wonderful to image other components in neuronal and non-neuronal cell types. However, as you said, that would take too long.
Is there some other channel that at low magnification (5X) provides as much independent information as a BIII tubulin channel in neurons? Probably not. The rich geometry of neuritic arbors can be a unique source of information as long as the monitored features show some independency. We already know that several features extracted by the Cellomics Neuronal Profiling Biotool (CNPB) are correlated not just for cells within a treatment but also among treatment population averages. We used this fact to reduce the dimensions of our data in Willie’s paper. However, in doing so, we may have missed some few treatments that were off-correlation. And this has two implications, one practical and one theoretical. The practical is that we may have missed hits. The theoretical is that it takes just one instance of an off-correlation treatment to break the absolute dependency status. We may then talk about levels of independency and even speculate that provided we are sampling a truly random portion of the target space, features with low independency score can be said to be determined by targets with low connectivity in the intracellular network in the conditions being evaluated. However, if we could measure all possible phenotypic features (Snapshots of transcriptome, proteome, metabolome and morphology. And their dynamics and higher level outcomes such as a behaviour in the case of an organism) would such a treatment be necessarily among the less perturbing ones?

More of the same on expanding phenotypic fingerprint at LemBix HCA, a context strongly biased toward neurons.

The level of phenotype similarity produced by any two treatments (gene overexpression, gene knock down compounds, etc) have been traditionally used as criteria to define if the targets in question operate through a similar mechanism.
Understandably, the power of any phenotype based association (is there some other type of association?) will be stronger with an increasing number of phenotypic features under consideration. As an extreme, but common case, there are many treatments that produce no apparent phenotype change even if they perturb the system. For them, it may be taking place changes in phenotypic features that are not being monitored. The so-called latent phenotypes.
Increasing the number of monitored phenotypic features can be used to study the specificity of action of a group of compounds prone to be rather unspecific: ATP-competitive protein kinase inhibitors. So, for example, for two molecules developed originally as inhibitors of PKX* (see below), even having the same scaffold, is the specificity profile the same? One way to access this issue is to test a panel of protein kinase to directly determine its specificity profile. However, there are at least two limitations in this otherwise very useful approach. One is that these panels are build with purified kinases in a test tube, conditions that differ from the intracellular environment. The second is that the molecule can have targetes other than kinases. Measuring the phenotypes these molecules produced in a multi-dimensional and meaningful phenotypic space is an important source of complementary and maybe more defining information. There are at least two ways to access this issue.

1-Use of few cell types that significantly differ in its molecular profiles. That is, for example, using neurons and hepatocytes to evaluate the effect of the compounds. There must be as many as possible trackable and independent (i.e. not correlated) phenotypic features as possible per cell type, in order to increase the resolving power. Neurons are particularly relevant in this sense. They are probably the cell type with the highest number of independent phenotypic features per channel.

2-Using a single cell type or a few similar cell types and trying combinations of drugs (epistasis) and culture conditions: The priming and the probe. The priming will alter the molecular landscape of the cell, producing or not, a different phenotype. Then under the influence of the primer, the probe is tested and the phenotype is recorded. In a sense, this is like using a different cell type because the primer perturbs the original type. Is there an strict demarcation line between cell types? Two probes that produce the same phenotype on the original cell type can produce different phenotypes in the primed conditions. This would be evidence that the probes don’t operate through the same mechanism. Instead of a priming drug, culture conditions can be also varied. The idea is to screen on a changed background network.

*Still, the most common model in use to generate new PK inhibitors (or of any other enzyme) is based on in vitro enzymatic or binding assay for the target but not necessarily working at the high-thoughput level but guided by rational design.